Research

An agent makes decisions before it answers: which team gets a ticket, whether a message holds personal data, whether to use a tool or ask a person.

The Orq.ai PII plugin, now in open beta, masks personal data before a request leaves the AI Gateway and restores it before the response reaches your user.

orq launch starts a coding agent already routed through the Orq.ai Gateway, with the workspace tools and skills borrowed for the session and nothing left on disk.

The Orq.ai CLI covers the entire Orq.ai API, and ships the Orq.ai Skills inside the binary so your coding agent knows how to use it.

Ranking evaluation data by how often an LLM judge flips its answer gives a review queue two to three times better than random sampling, with one blind spot.

Redacting PII from a RAG query costs a few points of recall for semantic search, but misses the right top document about 60% of the time for person queries.

What 5,624 adversarial attacks across Evaluatorq, PromptFoo, and DeepTeam tell us about attack quality, judge calibration.

Your eval pipeline has a judge. That judge has biases. Adding two more judges fixes that, but only if they disagree on the right things.

The cheapest accuracy upgrade in LLM engineering still works on GPT-5.4, Claude Sonnet 4.6, and Gemini 3 Pro. Here is what 334 runs taught us.

Every team building an LLM-as-a-judge pipeline faces the same question: does it matter where you put your instructions?

We explored using prompt optimization to make cheaper, smaller language models perform as well as expensive, more capable ones.

Key takeways from fine-tuning 8+ language models on a text classification task, from tiny 0.6B models to 14B behemoths.